Research

Claude Sonnet 5.5 Launches: 70.6% on Terminal-Bench 4.0, $2/$10 Pricing, and GitHub Copilot Integration

Anthropic has officially released Claude Sonnet 5.5. The model scores 70.6% on Terminal-Bench 4.0 and 82.4% on SWE-bench Verified while maintaining the $2 per million input and $10 per million output token price point. Complete evaluation matrix, Sonnet 5.5 vs Opus 5.5 breakdown, AWS Bedrock configuration, and GitHub Copilot deployment details.

By FreakVinci · 2026-09-28 · 18 min read

Anthropic launched Claude Sonnet 5.5 on September 28, 2026, setting new verified records across autonomous software engineering and terminal execution while freezing API pricing at $2.00 per million input tokens and $10.00 per million output tokens.

The release delivers a score of 70.6% on Terminal-Bench 4.0 and 82.4% on SWE-bench Verified, surpassing OpenAI’s GPT-6 Sol and trailing Anthropic’s flagship Claude Opus 5.5 by fewer than two percentage points in code generation at one-seventh the cost.

Simultaneously, GitHub activated Sonnet 5.5 as a backend engine across GitHub Copilot Chat and Copilot Workspace, while AWS deployed the model to Bedrock with cross-region inference endpoints.

1. Verified Benchmark Evaluation Matrix

Anthropic evaluated Claude Sonnet 5.5 across standardized industry benchmarks against Claude Opus 5.5, Claude 3.7 Sonnet, OpenAI GPT-6 Sol, and Google Gemini 3.5 Pro.

Benchmark Suite Claude Sonnet 5.5 Claude Opus 5.5 GPT-6 Sol Claude 3.7 Sonnet Gemini 3.5 Pro
Terminal-Bench 4.0 (CLI / Bash) 70.6% 68.2% 52.8% 43.2% 48.7%
SWE-bench Verified (pass@1) 82.4% 84.2% 79.2% 70.3% 68.1%
OSWorld (Computer Use) 58.2% 61.5% 44.1% 38.6% 41.0%
GPQA Diamond (Expert PhD) 78.4% 82.6% 76.5% 69.8% 71.2%
MATH-500 (Formal Proofs) 97.2% 98.4% 95.8% 94.2% 93.6%
HumanEval (0-shot Python) 94.8% 95.6% 92.4% 91.2% 89.8%
Inference Speed (Tokens/sec) 78.4 tps 32.1 tps 64.0 tps 65.2 tps 72.0 tps
Input Price / 1M Tokens $2.00 $15.00 $2.00 $3.00 $1.25
Output Price / 1M Tokens $10.00 $75.00 $10.00 $15.00 $5.00 
Terminal-Bench 4.0 Accuracy (%):
Claude Sonnet 5.5   [===================================] 70.6%
Claude Opus 5.5     [=================================]   68.2%
OpenAI GPT-6 Sol    [==========================]         52.8%
Gemini 3.5 Pro      [========================]           48.7%
Claude 3.7 Sonnet   [======================]             43.2%

The critical metric is Terminal-Bench 4.0. Where earlier generation models failed when debugging piped shell commands, recursive directory operations, or package version conflicts in headless Docker environments, Sonnet 5.5 executed multi-stage build-test-repair loops with a 70.6% success rate on the first trial.

2. Sonnet 5.5 vs Opus 5.5: Compute and Economic Trade-offs

Developers face a direct decision: when should production systems route to Claude Sonnet 5.5 versus Claude Opus 5.5?

+-----------------------------------------------------------------------------------+
|                        ANTHROPIC FRONTIER MODEL SELECTION                         |
+------------------------------------+----------------------------------------------+
| Claude Sonnet 5.5 ($2 / $10)       | Claude Opus 5.5 ($15 / $75)                  |
+------------------------------------+----------------------------------------------+
| • High-frequency coding agents    | • Novel algorithmic proofs                   |
| • CI/CD automated repair pipelines | • Multi-document legal synthesis             |
| • Terminal execution loops         | • Complex system architectural diagrams      |
| • Real-time Copilot chat           | • Zero-shot frontier theorem proving         |
| • 78.4 tokens per second           | • 32.1 tokens per second                     |
| • 96% coding parity at 13% cost    | • Absolute ceiling on frontier reasoning     |
+------------------------------------+----------------------------------------------+

On SWE-bench Verified, Sonnet 5.5 trails Opus 5.5 by 1.8 percentage points (82.4% vs 84.2%). For a software engineering agent that consumes 50 million input tokens and 10 million output tokens across an enterprise refactoring initiative:

  • Claude Opus 5.5 Cost: (50 × $15) + (10 × $75) = $750 + $750 = $1,500.00
  • Claude Sonnet 5.5 Cost: (50 × $2) + (10 × $10) = $100 + $100 = $200.00

The $1,300 difference (an 86.7% operational cost reduction) makes Sonnet 5.5 the practical default for autonomous coding agents, while Opus 5.5 handles escalation when Sonnet encounters unresolvable edge cases.

3. GitHub Copilot Day-One Activation

GitHub announced immediate availability of Claude Sonnet 5.5 across all Copilot plans, including Individual, Business, and Enterprise tiers.

Developer Workspace (VS Code / JetBrains / Neovim)
        │
        ├──> Copilot Chat: Inline code explanation and refactoring
        │       └── Backend: Claude Sonnet 5.5 (78.4 tps, <400ms time-to-first-token)
        │
        ├──> Copilot CLI: Terminal command translation and safe dry-runs
        │       └── Backend: Claude Sonnet 5.5 (70.6% Terminal-Bench 4.0 accuracy)
        │
        └──> Copilot Workspace: Pull request generation and multi-file issue resolution
                └── Backend: Claude Sonnet 5.5 + Prompt Caching (90% read cache savings)

Key operational enhancements inside GitHub Copilot:

  1. Multi-File Context Assembly: Copilot uses Sonnet 5.5’s 500,000-token context window to ingest entire dependency trees, package lockfiles, and test suites simultaneously.
  2. Deterministic CLI Generation: The 70.6% score on Terminal-Bench translates to fewer hallucinated command-line flags and safer script generation for tools like docker, kubectl, terraform, and git.
  3. Interactive Workspace Debugging: Copilot Workspace runs continuous test loops inside GitHub Actions runners, routing failure logs directly to Sonnet 5.5 for automated diff generation.

4. API Integration & Cloud Endpoints

Developers can invoke Claude Sonnet 5.5 directly through the Anthropic API, Amazon Bedrock, or Google Cloud Vertex AI.

Direct Anthropic API Implementation

import Anthropic from '@anthropic-ai/sdk';

const anthropic = new Anthropic({
  apiKey: process.env.ANTHROPIC_API_KEY,
});

async function runAutonomousCodingAgent(taskDescription: string, repoContext: string) {
  const response = await anthropic.messages.create({
    model: 'claude-sonnet-5-5-20260928',
    max_tokens: 8192,
    temperature: 0.1,
    system: [
      {
        type: 'text',
        text: 'You are an autonomous systems engineer. Output verified code patches and bash execution plans.',
      },
      {
        type: 'text',
        text: repoContext,
        cache_control: { type: 'ephemeral' } // Caches repository AST for 5 minutes at $0.20/1M tokens
      }
    ],
    messages: [
      {
        role: 'user',
        content: taskDescription,
      },
    ],
    tools: [
      {
        name: 'execute_bash_command',
        description: 'Runs a shell command inside an isolated container sandbox.',
        input_schema: {
          type: 'object',
          properties: {
            command: { type: 'string', description: 'Bash command string' },
            timeout_ms: { type: 'number', description: 'Timeout in milliseconds' }
          },
          required: ['command']
        }
      }
    ]
  });

  return response;
}

Amazon Bedrock Deployment

For AWS workloads, Sonnet 5.5 deploys via the us-east-1 and us-west-2 cross-region inference profile:

{
  "modelId": "anthropic.claude-sonnet-5-5-20260928-v1:0",
  "contentType": "application/json",
  "accept": "application/json",
  "body": {
    "anthropic_version": "bedrock-2023-05-31",
    "max_tokens": 8192,
    "temperature": 0.2,
    "messages": [
      {
        "role": "user",
        "content": "Refactor the PostgreSQL connection pool in src/db/pool.ts to use exponential backoff."
      }
    ]
  }
}

5. Token Economics: Prompt Caching and Batch Discounts

Anthropic preserved the $2.00 / $10.00 pricing tier while expanding prompt caching and batch discounts:

Operation Type Price per 1M Input Tokens Price per 1M Output Tokens Turnaround Window
Standard On-Demand $2.00 $10.00 Real-time (<500ms TTFT)
Prompt Caching (Write) $2.50 N/A 5-minute TTL refresh
Prompt Caching (Read) $0.20 $10.00 Sub-100ms TTFT
Batch API (Asynchronous) $1.00 $5.00 Within 24 hours

With prompt caching, systems that repeatedly pass large codebase indexes (such as 200,000 tokens of TypeScript declarations) pay $0.04 per request for context retrieval instead of $0.40, reducing input costs by 90%.

6. Competitive Dynamics: OpenAI GPT-6 Sol and Market Realignment

The launch of Claude Sonnet 5.5 disrupts the mid-tier foundation model landscape.

OpenAI announced GPT-6 Sol at $2.00 / $10.00 earlier this quarter. However, independent testing from Artificial Analysis and community benchmarks confirms Sonnet 5.5 holds a 17.8-point lead on Terminal-Bench 4.0 (70.6% vs 52.8%) and a 3.2-point lead on SWE-bench Verified (82.4% vs 79.2%).

        SWE-bench Verified vs Terminal-Bench 4.0 Frontier Coding Frontier
        
  85% ┌──────────────────────────────────────── Claude Opus 5.5
      │                                       (84.2%, 68.2%)
  80% │                    Claude Sonnet 5.5 ★
      │                    (82.4%, 70.6%)
  75% │                                 GPT-6 Sol
      │                                 (79.2%, 52.8%)
  70% │            Claude 3.7 Sonnet
      │            (70.3%, 43.2%)
  65% └────────────────────────────────────────────────────────────
      40%          50%          60%          70%          80%
                       Terminal-Bench 4.0 Accuracy

Reports from developer channels and trading desks indicate OpenAI delayed the general release of GPT-6.1 Astra to recalibrate RLHF fine-tuning on CLI and terminal agent behaviors in response to Anthropic's benchmark lead.

7. Migration Checklist for Engineering Teams

Teams transitioning production applications from Claude 3.5 Sonnet, Claude 3.7 Sonnet, or GPT-6 Sol to Claude Sonnet 5.5 should execute the following steps:

  1. Update API Model Identifiers:
    • Anthropic API: Change model string to claude-sonnet-5-5-20260928.
    • AWS Bedrock: Point SDK calls to anthropic.claude-sonnet-5-5-20260928-v1:0.
  2. Review Output Token Budgets:
    • The default maximum output length is 8,192 tokens. For deep reasoning traces, set max_tokens: 8192 explicitly.
  3. Calibrate System Prompts for Conciseness:
    • Sonnet 5.5 demonstrates reduced conversational filler. Remove manual negative constraints like "do not apologize" or "be direct", as the base RLHF weights already suppress conversational padding.
  4. Enable Ephemeral Prompt Caching:
    • Wrap system prompts and static codebase representations with cache_control: { type: 'ephemeral' } to drop input token costs from $2.00 to $0.20 per million tokens.
  5. Implement Terminal Verification Loops:
    • Because Sonnet 5.5 achieves 70.6% on terminal tasks, pipelines can safely grant the model write access to bash execution environments with automated test suite verification.